Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * linux/arch/m32r/lib/memset.S |
| 3 | * |
| 4 | * Copyright (C) 2001,2002 Hiroyuki Kondo, and Hirokazu Takata |
| 5 | * Copyright (C) 2004 Hirokazu Takata |
| 6 | * |
| 7 | * void *memset(void *dst, int val, int len); |
| 8 | * |
| 9 | * dst: r0 |
| 10 | * val: r1 |
| 11 | * len: r2 |
| 12 | * ret: r0 |
| 13 | * |
| 14 | */ |
| 15 | /* $Id$ */ |
| 16 | |
| 17 | #include <linux/config.h> |
| 18 | |
| 19 | .text |
| 20 | .global memset |
| 21 | |
| 22 | #ifdef CONFIG_ISA_DUAL_ISSUE |
| 23 | |
| 24 | .align 4 |
| 25 | memset: |
| 26 | mv r4, r0 || cmpz r2 |
| 27 | jc r14 |
| 28 | cmpui r2, #16 |
| 29 | bnc qword_align_check |
| 30 | cmpui r2, #4 |
| 31 | bc byte_set |
| 32 | word_align_check: /* len >= 4 */ |
| 33 | and3 r3, r4, #3 |
| 34 | beqz r3, word_set |
| 35 | addi r3, #-4 |
| 36 | neg r3, r3 /* r3 = -(r3 - 4) */ |
| 37 | align_word: |
| 38 | stb r1, @r4 || addi r4, #1 |
| 39 | addi r2, #-1 || addi r3, #-1 |
| 40 | bnez r3, align_word |
| 41 | cmpui r2, #4 |
| 42 | bc byte_set |
| 43 | word_set: |
| 44 | and3 r1, r1, #0x00ff /* r1: abababab <-- ??????ab */ |
| 45 | sll3 r3, r1, #8 |
| 46 | or r1, r3 || addi r4, #-4 |
| 47 | sll3 r3, r1, #16 |
| 48 | or r1, r3 || addi r2, #-4 |
| 49 | word_set_loop: |
| 50 | st r1, @+r4 || addi r2, #-4 |
| 51 | bgtz r2, word_set_loop |
| 52 | bnez r2, byte_set_wrap |
| 53 | st r1, @+r4 |
| 54 | jmp r14 |
| 55 | |
| 56 | qword_align_check: /* len >= 16 */ |
| 57 | and3 r3, r4, #15 |
| 58 | bnez r3, word_align_check |
| 59 | qword_set: |
| 60 | and3 r1, r1, #0x00ff /* r1: abababab <-- ??????ab */ |
| 61 | sll3 r3, r1, #8 |
| 62 | or r1, r3 || addi r4, #-4 |
| 63 | sll3 r3, r1, #16 |
| 64 | or r1, r3 || ldi r5, #16 |
| 65 | qword_set_loop: |
| 66 | ld r3, @(4,r4) /* cache line allocate */ |
| 67 | st r1, @+r4 || addi r2, #-16 |
| 68 | st r1, @+r4 || cmpu r2, r5 |
| 69 | st r1, @+r4 |
| 70 | st r1, @+r4 |
| 71 | bnc qword_set_loop || cmpz r2 |
| 72 | jc r14 |
| 73 | set_remainder: |
| 74 | cmpui r2, #4 |
| 75 | bc byte_set_wrap1 |
| 76 | addi r2, #-4 |
| 77 | bra word_set_loop |
| 78 | |
| 79 | byte_set_wrap: |
| 80 | addi r2, #4 |
| 81 | cmpz r2 |
| 82 | jc r14 |
| 83 | byte_set_wrap1: |
| 84 | addi r4, #4 |
| 85 | #if defined(CONFIG_ISA_M32R2) |
| 86 | byte_set: |
| 87 | addi r2, #-1 || stb r1, @r4+ |
| 88 | bnez r2, byte_set |
| 89 | #elif defined(CONFIG_ISA_M32R) |
| 90 | byte_set: |
| 91 | addi r2, #-1 || stb r1, @r4 |
| 92 | addi r4, #1 |
| 93 | bnez r2, byte_set |
| 94 | #else |
| 95 | #error unknown isa configuration |
| 96 | #endif |
| 97 | end_memset: |
| 98 | jmp r14 |
| 99 | |
| 100 | #else /* not CONFIG_ISA_DUAL_ISSUE */ |
| 101 | |
| 102 | .align 4 |
| 103 | memset: |
| 104 | mv r4, r0 |
| 105 | beqz r2, end_memset |
| 106 | cmpui r2, #16 |
| 107 | bnc qword_align_check |
| 108 | cmpui r2, #4 |
| 109 | bc byte_set |
| 110 | word_align_check: /* len >= 4 */ |
| 111 | and3 r3, r4, #3 |
| 112 | beqz r3, word_set |
| 113 | addi r3, #-4 |
| 114 | neg r3, r3 /* r3 = -(r3 - 4) */ |
| 115 | align_word: |
| 116 | stb r1, @r4 |
| 117 | addi r4, #1 |
| 118 | addi r2, #-1 |
| 119 | addi r3, #-1 |
| 120 | bnez r3, align_word |
| 121 | cmpui r2, #4 |
| 122 | bc byte_set |
| 123 | word_set: |
| 124 | and3 r1, r1, #0x00ff /* r1: abababab <-- ??????ab */ |
| 125 | sll3 r3, r1, #8 |
| 126 | or r1, r3 |
| 127 | sll3 r3, r1, #16 |
| 128 | or r1, r3 |
| 129 | addi r2, #-4 |
| 130 | addi r4, #-4 |
| 131 | word_set_loop: |
| 132 | st r1, @+r4 |
| 133 | addi r2, #-4 |
| 134 | bgtz r2, word_set_loop |
| 135 | bnez r2, byte_set_wrap |
| 136 | st r1, @+r4 |
| 137 | jmp r14 |
| 138 | |
| 139 | qword_align_check: /* len >= 16 */ |
| 140 | and3 r3, r4, #15 |
| 141 | bnez r3, word_align_check |
| 142 | qword_set: |
| 143 | and3 r1, r1, #0x00ff /* r1: abababab <-- ??????ab */ |
| 144 | sll3 r3, r1, #8 |
| 145 | or r1, r3 |
| 146 | sll3 r3, r1, #16 |
| 147 | or r1, r3 |
| 148 | addi r4, #-4 |
| 149 | qword_set_loop: |
| 150 | ld r3, @(4,r4) /* cache line allocate */ |
| 151 | addi r2, #-16 |
| 152 | st r1, @+r4 |
| 153 | st r1, @+r4 |
| 154 | cmpui r2, #16 |
| 155 | st r1, @+r4 |
| 156 | st r1, @+r4 |
| 157 | bnc qword_set_loop |
| 158 | bnez r2, set_remainder |
| 159 | jmp r14 |
| 160 | set_remainder: |
| 161 | cmpui r2, #4 |
| 162 | bc byte_set_wrap1 |
| 163 | addi r2, #-4 |
| 164 | bra word_set_loop |
| 165 | |
| 166 | byte_set_wrap: |
| 167 | addi r2, #4 |
| 168 | beqz r2, end_memset |
| 169 | byte_set_wrap1: |
| 170 | addi r4, #4 |
| 171 | byte_set: |
| 172 | addi r2, #-1 |
| 173 | stb r1, @r4 |
| 174 | addi r4, #1 |
| 175 | bnez r2, byte_set |
| 176 | end_memset: |
| 177 | jmp r14 |
| 178 | |
| 179 | #endif /* not CONFIG_ISA_DUAL_ISSUE */ |
| 180 | |
| 181 | .end |